Development of a neuro-fuzzy approach for treating multimodal distributions and spurious clusters

In this paper, we introduce a developed approach of neuro-fuzzy model free estimators to treat decision problems, where presence of highly multimodal distribution and spurious clusters in decision space causes problems in estimation. After critical analysis of Bart Kosko's DCL – AVQ + FAM neuro-fuzzy system, we developed our Straitjacket + F-wing system. Straitjacket neural system is an improved version of Kohonen's feature mapping according to convergence and termination conditions. F-wing fuzzy system bases on the feature map produced by Straitjacket and provides faster and more exact estimator interface than Bart Kosko's FAM, using fuzzy wing functions instead of fuzzy hyperpyramids. At the end of the paper we perform a test on artificial database comparing the efficiency of our approach with Kosko's system, discriminant analysis, hierarchic and K-mean clustering.